Progress Update
The project began from an interest in cyber-resilient connected and automated vehicles, specifically longitudinal vehicle platooning under unreliable or adversarial vehicle-to-vehicle communication. The first working idea was to combine physics-based consistency checking with a graph model that could assign trust scores to received V2V messages.
The initial research question was roughly: can a platoon infer whether a communicated vehicle state is trustworthy by comparing it against local dynamics, neighboring vehicles, and the communication graph? This was initially close to an attack-detection project. The key conceptual shift was to ask a harder and more useful question: even if a suspicious message can be detected, how should that trust estimate change the control action, and what prevents a mistaken trust estimate from causing a collision?
Technical Essay
The central idea emerged as a four-layer control architecture:
- Treat every received V2V packet as a directed edge claim from sender j to receiver i.
- Estimate edge-level reliability from communicated state, receiver state, relative motion, physical consistency, statistical consistency, and message freshness.
- Insert the estimated trust weights directly into cooperative adaptive cruise control rather than treating attack detection as a disconnected classifier.
- Supervise the final longitudinal command with a runtime gap/TTC shield so that learned trust is never treated as a safety certificate.
This architecture separated two jobs that are frequently conflated. The trust model performs information selection; the shield performs last-mile safety intervention. That separation became the defining idea of SafeGraph-CAV.
Reading Note
Early reading centered on three bodies of work: classical CACC and string stability, cyberattacks against V2V-enabled platoons, and graph learning for relational systems. The literature made the gap visible. Existing secure-control methods addressed spoofing, false-data injection, denial of service, or event-triggered control; graph learning offered relational inductive bias; runtime safety methods could override unsafe actions. However, these pieces were often studied separately. The project would connect relational trust, control integration, and runtime shielding in one empirical system.